Enabling Search for Facts and
Implied Facts in Historical Documents
David W. Embley, Stephen W. Liddle, Deryle W.
Lonsdale, Spencer Machado, Thomas Packer,
Joseph Park, Nathan Tate, Andrew Zitzelberger
Brigham Young University
BYU Data Extraction Research Group
WoK-HD
(A Web of Knowledge Superimposed over Historical Documents)
…
…
…
…
WoK-HD
(A Web of Knowledge Superimposed over Historical Documents)
grandchildren of Mary Ely
…
…
…
…
WoK-HD
(A Web of Knowledge Superimposed over Historical Documents)
grandchildren of Mary Ely
…
…
…
…
WoK-HD
(A Web of Knowledge Superimposed over Historical Documents)
grandchildren of Mary Ely
…
…
…
…
WoK-HD
(A Web of Knowledge Superimposed over Historical Documents)
grandchildren of Mary Ely
…
…
…
…
WoK-HD Input
Querying for Facts & Implied Facts
Querying for Facts & Implied Facts
Animation of
1. Extraction query, results, highlighting
2. Reasoned Query, results, reasoning chain, highlighting
Extraction Ontologies
Extraction Ontologies
Fact Extraction
Fact Extraction
Fact Extraction
Reasoning for Implied Facts
Reasoning for Implied Facts
Reasoning for Implied Facts
Reasoning for Implied Facts
Query Interpretation
“Mary Ely” grandchild
Query Interpretation
“Mary Ely” grandchild
Query Interpretation
“Mary Ely” grandchild
Generated SPARQL Query
Generated SPARQL Query
Query Results
Results of Processing
the Ely Ancestry (all 830 Pages)
• Number of facts extracted: 22,251
– 8,740 Person-Birthdate facts
– 3,803 Person-Deathdate facts
– 9,708 children facts, including
• 5,020 Child-has-parent-Person facts
• 2,394 Son-of-Person facts
• 2,294 Daughter-of-Person facts
• Number of implied grandchild facts inferred: 5,277
• Processing time:
– ~18 seconds per page
– CPU time: ~4 hours
– Processing 10 in parallel: ~24 minutes
Results of Processing
the Ely Ancestry (all 830 Pages)
• Precision: .52
(by randomly selecting & checking 100 of the 22,251 facts)
• Recall: .33 & Precision: .40
(by randomly selecting and checking 2 fact-filled family pages)
• Errors:
– Name recognizer
– Text pattern expectations
– OCR
• Varying accuracy (for pages checked)
–
–
–
–
Recall: .11, Precision: .11 (bad combination of all problems)
Recall: .50, Precision: .68 (some problems, but closer to expectations)
Recall: .59, Precision: .71 (10 pages, mostly as expected)
Recall: .91, Precision: .94 (tuned, no problems except a few OCR errors)
Current and Future Work
• Implementation Status:
– Full line works (but is fragile & needs finishing touches)
– HyKSS integrated (but not all features)
• Scalability:
– Handcrafted extraction ontologies & reasoning rules
(worth the work for certain applications)
– ListReader (plus bootstrapping for lists and general extraction)
– Optimization (especially for query processing)
• Integration:
– Mapping extraction ontologies to domain ontologies
– Object identity for people and places
Summary and Conclusion
• WoK-HD
– Superimposes a web of knowledge over a
collection of historical documents
– Works as a proof-of-concept prototype
• To build and deploy the WoK-HD successfully:
– Efficient implementation
– Better, more cost-effective extraction
– Integration and record linkage
www.deg.byu.edu